A robust incremental low-light image enhancement method and device

Abstract

The present invention discloses a robust incremental low-light image enhancement method and device. The present invention collects a series of low-light image data sets and corresponding normal-light images to construct an incremental sample library; constructs a domain incremental low-light image enhancement model; obtains the sample domain that needs to be trained in accordance with the training order, and preprocesses the sample domain in real time; uses the preprocessed training data set to train the domain incremental low-light image enhancement model; after training, uses the current sample domain verification set to filter out the domain incremental low-light image enhancement model with the highest evaluation index and saves it; then, uses all old domains and current domain test sets to comprehensively test the model, evaluates the enhancement effect and forgetting rate, and uses the current model parameters to update the model parameters of the next incremental stage; completes the training for all domains, and obtains the final domain incremental low-light image enhancement model. The present invention adaptively enhances the dynamically changing low-light image domain, and the enhanced image effect is better than the prior art.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and in particular to an incremental low-light image enhancement method and device. Background Art

[0002] With the continuous development of computer vision, low-light image enhancement has become one of the key research directions and is widely used in fields such as autonomous driving, intelligent monitoring, image segmentation, and image classification. In recent years, low-light image enhancement methods based on deep learning have made great progress. However, this excellent performance is achieved under the premise that the input data has extremely similar or identical distribution. Existing models cannot effectively enhance low-light images with large differences. This is mainly due to the differences in the environment in which the low-light images are taken and the differences in hardware devices. The scenes, brightness, and noise of low-light images are variable. This variability makes it impossible for existing models to adaptively learn dynamically changing low-light images. Therefore, it is a challenging task to effectively restore the image quality of low-light images captured under various environmental conditions.

[0003] Incremental learning is an effective way to solve the above problems. Incremental learning refers to a learning method that continuously acquires new knowledge in a new environment while retaining old knowledge. It is widely used in advanced visual fields such as autonomous driving, image classification and target detection. However, low-vision fields such as low-light image enhancement have not been explored so far. Therefore, low-light image enhancement based on incremental learning can gradually learn various data distributions in the low-light data domain and realize effective information recovery of various colors, details, textures, etc. In addition, this invention can lay the foundation for performance improvement in many fields, such as autonomous driving, security monitoring and medical imaging. Summary of the invention

[0004] Purpose of the invention: In view of the problems existing in the prior art, the present invention provides a robust incremental low-light image enhancement method and device.

[0005] Technical solution: The robust incremental low-light image enhancement method of the present invention comprises:

[0006] Step S1: construct an incremental sample library, collect a series of low-light image datasets and corresponding normal-light images, each domain dataset includes a training dataset, a validation dataset and a test dataset, and determine the training order of different domain datasets in the incremental sample library;

[0007] Step S2: construct a domain-incremental low-light image enhancement model to adapt to the low-light image enhancement requirements of different contexts; the model includes a feature attention consistency module and an adversarial perturbation module;

[0008] Step S3: Obtain the sample domain that needs to be trained currently according to the training sequence, and pre-process the sample domain in real time;

[0009] Step S4: Using the preprocessed training data set, the domain incremental low-light image enhancement model is efficiently trained. If the current training stage is the base training stage, the current sample domain is input into the model; if the current training stage is the incremental training stage, the current sample domain is input into the old domain model frozen in the previous incremental stage and the new domain model in the current stage respectively, and then the images output by the two models are input into the dynamic discriminator network for adversarial learning;

[0010] Step S5: Use the current sample domain verification set to select the domain increment low-light image enhancement model with the highest evaluation index and save it;

[0011] Step S6: Use all old domain and current domain test sets to comprehensively test the model, evaluate the enhancement effect and forgetting rate, and use the current model parameters to update the model parameters of the next incremental stage;

[0012] Step S7: Return to step S3 and execute until the training for all domains is completed to obtain the final efficiently trained domain incremental low-light image enhancement model.

[0013] Furthermore, in step S1, the training order of different domain data sets in the incremental sample library is determined as follows:

[0014] In the incremental learning setting, a domain dataset is randomly selected from the incremental sample library as the base dataset for the first training, and the remaining domain datasets are used as incremental domains. A single domain dataset is sequentially selected from the remaining incremental sample library as the domain dataset for the next training.

[0015] Furthermore, in step S2, the domain incremental low-light image enhancement model includes a feature attention consistency module and an adversarial perturbation module. The feature attention consistency module includes a learnable attention that interacts with the hidden layer features of each scale, focuses on the important parts of the current domain features, and constrains the learnable attention of the previous incremental stage to the learnable attention of the current stage, retaining fine-grained feature knowledge; the adversarial perturbation module includes a learnable perturbation that is superimposed on the output of the current incremental stage model, and performs adversarial learning on the output of the previous stage model and the output of the current incremental stage model, retaining coarse-grained representation knowledge;

[0016] Furthermore, in step S3, the method for preprocessing the sample domain in real time is specifically as follows:

[0017] Step S3-1: randomly flipping the low-light image horizontally and vertically with a preset probability;

[0018] Step S3-2: Crop the low-light image to a uniform size;

[0019] Step S3-3: convert the low-light image into a tensor, and convert the data type into float;

[0020] Step S3-4: normalize each pixel of the low-light image so that the pixel value falls within the range of [-1, 1].

[0021] Furthermore, in step S4, using the preprocessed training data set to efficiently train the domain incremental low-light image enhancement model includes:

[0022] Step S4-1: Setting training parameters. If the current stage is the base training stage, the domain incremental low-light image enhancement model is randomly initialized. If it is the incremental training stage, the network parameters obtained by the previous incremental stage training are directly used for initialization.

[0023] Step S4-2: In each training phase, randomly initialize a learnable attention and a learnable perturbation;

[0024] Step S4-3: In the base training phase, the preprocessed training set is input into the current domain model for training, and the trained model parameters are saved. In the incremental training phase, the samples of the same batch are input into the frozen old domain model and the new domain model for training;

[0025] Step S4-4: In the base training stage, multi-scale features are obtained through the feature extractor, and the first set of operations are first performed on each feature map: first, a pooling operation is performed to obtain low-frequency features, and then the previous layer features are multiplied by the current features to interact between features, and then the obtained features are convolved, activated, and dimensionally expanded to obtain significant features, which are then multiplied by the learnable attention; in the incremental training stage, the first set of operations is the same as the first set of operations in the base training stage;

[0026] Step S4-5: In the base training stage, a second set of operations is performed: the attention features obtained by the first set of operations are averaged, convolved, and activated, and then the reduced dimension features are superimposed on the next layer; in the incremental training stage, the second set of operations is the same as the second set of operations in the base training stage, and in addition, the feature attention consistency loss is calculated using the learnable attention in the current stage and the frozen attention in the previous stage;

[0027] Step S4-6: In the base training stage, only the original loss function needs to be calculated for the obtained output image; in the incremental training stage, a learnable perturbation is superimposed on the visual output image of the current stage model, and then the perturbation-superimposed image and the visual image output by the model in the previous incremental stage are respectively input into the dynamic discriminator network for adversarial learning, and then the output result of the discriminator is used to calculate the relative adversarial loss;

[0028] Step S4-7: Combine the losses calculated in steps S4-5 and S4-6 to optimize the entire network parameters.

[0029] Furthermore, in step S4-5, the feature attention consistency loss is calculated using the learnable attention in the current stage and the frozen attention in the previous stage, which is specifically implemented as follows:

[0030] First, the low-light training dataset of the current incremental stage t is in Represents low-light images, represents the corresponding normal light image, N t Represents the number of samples in the current domain, Input feature extractor separately and a frozen feature extractor The following formula is used to extract the intermediate hidden layer features:

[0031]

[0032] In the formula, L represents the number of feature hidden layers; indicates enhanced low-light pictures; represents the i-th hidden layer, i belongs to 1,...,L-1,L; refers to a nested function where The corresponding hidden layer features obtained by the feature extractor are: in

[0033] In order to learn the important features of the new domain and retain the important features of the old domain, and maintain the balance between stability and plasticity, a learnable attention A is introduced for each incremental stage. The current incremental stage attention is A t , the attention A of the previous incremental stage t-1 Guide the learning of attention in the current incremental stage, retain the important feature knowledge of the old domain, and use the feature attention consistency loss based on mean square error to optimize the model. The formula is as follows:

[0034]

[0035] In the formula, represents the feature attention consistency loss at the t-th incremental stage.

[0036] Furthermore, in step S4-6, the relative adversarial loss is calculated using the output result of the discriminator, and the specific implementation method is as follows:

[0037] First, the visual output of the old domain model is defined as the real sample, the visual output of the new domain model is defined as the generated sample, and a learnable perturbation R is superimposed on the visual output of the new domain model. t , designed to damage the image quality; then the generated samples and real samples with superimposed perturbations are input into the dynamic discriminator, and the output of the old domain model frozen in the previous incremental stage is expressed as: The output of the new domain model is represented as: Conduct adversarial training;

[0038] In the incremental stage t, to optimize the generator, that is, the feature extractor The relative adversarial loss based on the least squares method is formulated as follows:

[0039]

[0040] In the formula, Represents the input data x t The expected value of p(x t ) represents the input data x t The true probability distribution of .

[0041] In the incremental stage t, to optimize the discriminator The relative adversarial loss based on the least squares method is formulated as follows:

[0042]

[0043] In the formula, the generator Try to minimize the objective function The adversarial discriminator Try to maximize it.

[0044] Furthermore, in step S4-7, in order to optimize the backbone network, i.e., the feature extractor Total loss The following formula:

[0045]

[0046] In the formula, represents the original loss function of the backbone network in the tth incremental stage, γ and δ are two hyperparameters, represents the feature attention consistency loss at the t-th incremental stage, represents the least squares-based relative adversarial loss at the incremental stage t.

[0047] To optimize the dynamics discriminator Network, total loss The following formula:

[0048]

[0049] Where η is a hyperparameter, is the optimized discriminator in the incremental stage t A least squares based relative adversarial loss is used.

[0050] Furthermore, in step S6, the model is comprehensively tested using a test data set to evaluate the enhancement effect and the forgetting rate. After the current domain model is trained, the model parameters obtained by the training are used to initialize the model parameters of the next incremental stage.

[0051] A robust incremental low-light image enhancement device described in the present invention includes a processor and a computer program stored in a memory and executable on the processor, and the processor implements the above method when executing the program.

[0052] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0053] (1) The submodule in the domain-incremental low-light image enhancement model of the present invention, the feature attention consistency module, focuses on the important features of the new domain while retaining the fine-grained feature knowledge of the old domain, alleviating domain shift.

[0054] (2) The submodule in the domain incremental low-light image enhancement model of the present invention, the adversarial perturbation module, performs adversarial learning on the visual output of the model by superimposing perturbations on the visual output of the old domain model, thereby retaining the coarse-grained representation knowledge of the old domain and enhancing the robustness of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a flowchart of the domain incremental low-light image enhancement method provided by the present invention;

[0056] Figure 2 It is a schematic diagram of the overall structure of the domain incremental low-light image enhancement model of the present invention.

[0057] Figure 3 It is a schematic diagram of the detailed structure of the submodules of the present invention. DETAILED DESCRIPTION

[0058] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiment is only one of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0059] This embodiment specifically provides a robust incremental low-light image enhancement method, such as Figure 1 and Figure 2 As shown, the following steps are included:

[0060] Step S1: construct an incremental sample library, collect a series of low-light image datasets and corresponding normal-light images, each domain dataset includes a training dataset, a validation dataset and a test dataset, and determine the training order of different domain datasets in the incremental sample library;

[0061] This embodiment uses the three data sets LOL-v1-RELLISUR-LOLv2 as the incremental sample library.

[0062] Step S2: construct a domain-incremental low-light image enhancement model to adapt to the low-light image enhancement requirements of different contexts;

[0063] like Figure 2 As shown, the domain incremental low-light image enhancement model includes a feature extractor (generator) and a dynamic discriminator. The feature extractor is used to extract features from the input samples, and can be a low-light image enhancement model in previous work, such as a model based on a convolutional neural network and a model based on a Transformer; the dynamic discriminator is a lightweight convolution-based four downsampling layers and a 1x1 convolution operation used to discriminate the real samples of the generated samples.

[0064] Step S3: Obtain the sample domain that needs to be trained currently according to the training sequence, and pre-process the sample domain in real time;

[0065] Among them, the specific method of preprocessing the sample domain in real time is as follows:

[0066] Step S3-1: randomly flipping the low-light image horizontally and vertically with a preset probability;

[0067] Step S3-2: Crop the low-light image to a uniform size;

[0068] Step S3-3: convert the low-light image into a tensor, and convert the data type into float;

[0069] Step S3-4: normalize each pixel of the low-light image so that the pixel value falls within the range of [-1, 1].

[0070] Step S4: training a domain incremental low-light image enhancement model using the preprocessed training dataset;

[0071] The training method is:

[0072] Step S4-1: Setting training parameters. If the current stage is the base training stage, the domain incremental low-light image enhancement model is randomly initialized. After training, its network parameters are saved. If it is the incremental training stage, the network parameters obtained by the previous incremental stage training are directly used for initialization. The training parameters include the number of iterations, learning rate and other necessary parameter settings.

[0073] Step S4-2: In each training phase, randomly initialize a learnable attention and a learnable perturbation;

[0074] Step S4-3: In the base training phase, the preprocessed training set is input into the current domain model for training, and the trained model parameters are saved. In the incremental training phase, the samples of the same batch are input into the frozen old domain model and the new domain model for training;

[0075] Steps S4-4 and S4-5 are specifically implemented as follows: First, the low-light training data set of the current incremental stage t is in Represents low-light images, represents the corresponding normal light image, N t Represents the number of samples in the current domain, Input feature extractor separately and a frozen feature extractor The following formula is used to extract the intermediate hidden layer features:

[0076]

[0077] In the formula, L represents the number of feature hidden layers; indicates enhanced low-light pictures; represents the i-th hidden layer, i belongs to 1,...,L-1,L; refers to a nested function where The corresponding hidden layer features obtained by the feature extractor are: in

[0078] A learnable attention A is introduced for each incremental stage, and the attention of the current incremental stage is A t , the attention A of the previous incremental stage t-1Guide the learning of attention in the current incremental stage, retain the important feature knowledge of the old domain, and use the feature attention consistency loss based on mean square error to optimize the model. The formula is as follows:

[0079]

[0080] In the formula, represents the feature attention consistency loss at the t-th incremental stage.

[0081] Step S4-6 is specifically implemented as follows: First, the visual output of the old domain model is defined as the real sample, and the visual output of the new domain model is defined as the generated sample. A learnable perturbation R is superimposed on the visual output of the new domain model. t , designed to damage the image quality; then the generated samples and real samples with superimposed perturbations are input into the dynamic discriminator, and the output of the old domain model frozen in the previous incremental stage is expressed as: The output of the new domain model is represented as: Conduct adversarial training;

[0082] In the incremental stage t, to optimize the generator, that is, the feature extractor The relative adversarial loss based on the least squares method is formulated as follows:

[0083]

[0084] In the formula, Represents the input data x t The expected value of p(x t ) represents the input data x t The true probability distribution of .

[0085] In the incremental stage t, to optimize the discriminator The relative adversarial loss based on the least squares method is formulated as follows:

[0086]

[0087] In the formula, the generator Try to minimize the objective function The adversarial discriminator Try to maximize it.

[0088] Step S4-7: Combine the losses calculated in steps S4-5 and S4-6 to optimize the network parameters of the entire domain low-light image enhancement model.

[0089] To optimize the backbone network, i.e. the feature extractor Total loss The following formula:

[0090]

[0091] In the formula, represents the original loss function of the backbone network in the tth incremental stage, γ and δ are two hyperparameters, represents the feature attention consistency loss at the t-th incremental stage, represents the least squares-based relative adversarial loss at the incremental stage t.

[0092] To optimize the dynamics discriminator Network, total loss The following formula:

[0093]

[0094] Where η is a hyperparameter, is the optimized discriminator in the incremental stage t A least squares based relative adversarial loss is used.

[0095] Step S5: Use the current sample domain verification set to select the domain increment low-light image enhancement model with the highest evaluation index and save it;

[0096] Step S6: Use all old domain and current domain test sets to comprehensively test the model, evaluate the enhancement effect and forgetting rate, and use the current model parameters to update the model parameters of the next incremental stage;

[0097] Step S7: Return to step S3 and execute until the training for all domains is completed, and the final domain incremental low-light image enhancement model is obtained.

[0098] The present invention discloses a robust incremental low-light image enhancement method and device. The present invention first collects a series of low-light image data sets and corresponding normal-light images to construct an incremental sample library; then constructs a domain incremental low-light image enhancement model to adapt to the low-light image enhancement needs of different contexts. The model includes a feature attention consistency module and an adversarial perturbation module. The feature attention consistency module includes a learnable attention that interacts with hidden layer features of each scale, focuses on important parts of the current domain features, and constrains the learnable attention of the previous incremental stage to the learnable attention of the current stage, retaining the fine-grained feature knowledge of the old domain; the adversarial perturbation module includes a learnable perturbation that is superimposed on the current incremental stage. The output of the model is to conduct adversarial learning between the output of the model in the previous stage and the output of the model in the current incremental stage, and retain the coarse-grained representation knowledge; then, the sample domain that needs to be trained is obtained in the training order, and the sample domain is preprocessed in real time; the domain incremental low-light image enhancement model is trained using the preprocessed training data set; after training, the domain incremental low-light image enhancement model with the highest evaluation index is selected using the current sample domain verification set and saved; then, the model is comprehensively tested using all old domains and current domain test sets to evaluate the enhancement effect and forgetting rate, and the current model parameters are used to update the model parameters of the next incremental stage; training for all domains is completed to obtain the final domain incremental low-light image enhancement model. The present invention can adaptively enhance the dynamically changing low-light image domain, and the enhanced image effect is better than the prior art.

Claims

1. A robust incremental low-light image enhancement method, characterized in that The method includes: Step S1: construct an incremental sample library, collect a series of low-light image datasets and corresponding normal-light images, each domain dataset includes a training dataset, a validation dataset and a test dataset, and determine the training order of different domain datasets in the incremental sample library; Step S2: construct a domain-incremental low-light image enhancement model to adapt to the low-light image enhancement requirements of different contexts. The model includes a feature attention consistency module and an adversarial perturbation module. Step S3: Obtain the sample domain that needs to be trained currently according to the training sequence, and pre-process the sample domain in real time; Step S4: using the preprocessed training data set to train the domain incremental low-light image enhancement model, if the current training stage is the base training stage, the current sample domain is input into the model; if the current training stage is the incremental training stage, the current sample domain is input into the old domain model frozen in the previous incremental stage and the new domain model in the current stage respectively, and then the images output by the two models are input into the dynamic discriminator network for adversarial learning; Step S5: Use the current sample domain verification set to select the domain increment low-light image enhancement model with the highest evaluation index and save it; Step S6: Use all old domain and current domain test sets to comprehensively test the model, evaluate the enhancement effect and forgetting rate, and use the current model parameters to update the model parameters of the next incremental stage; Step S7: Return to step S3 and execute until the training for all domains is completed to obtain the final domain increment low-light image enhancement model; In step S2, the domain incremental low-light image enhancement model includes a feature attention consistency module and an adversarial perturbation module, wherein the feature attention consistency module focuses on important parts of the current domain features and retains fine-grained feature knowledge of the old domain; and the adversarial perturbation module retains coarse-grained representation knowledge of the old domain; In step S4, using the preprocessed training data set to train the domain incremental low-light image enhancement model includes: Step S4-1: Setting training parameters. If the current stage is the base training stage, the domain incremental low-light image enhancement model is randomly initialized. If it is the incremental training stage, the network parameters obtained by the previous incremental stage training are directly used for initialization. Step S4-2: In each training phase, randomly initialize a learnable attention and a learnable perturbation; Step S4-3: In the base training phase, the preprocessed training set is input into the current domain model for training, and the trained model parameters are saved. In the incremental training phase, the samples of the same batch are input into the frozen old domain model and the new domain model for training; Step S4-4: In the base training stage, multi-scale features are obtained through the feature extractor, and the first set of operations are first performed on each feature map: first, a pooling operation is performed to obtain low-frequency features, and then the previous layer features are multiplied by the current features to interact between features, and then the obtained features are convolved, activated, and dimensionally expanded to obtain significant features, which are then multiplied by the learnable attention; in the incremental training stage, the first set of operations is the same as the first set of operations in the base training stage; Step S4-5: In the base training stage, a second set of operations is performed: the attention features obtained by the first set of operations are averaged, convolved, and activated, and then the reduced dimension features are superimposed on the next layer; in the incremental training stage, the second set of operations is the same as the second set of operations in the base training stage, and in addition, the feature attention consistency loss is calculated using the learnable attention in the current stage and the frozen attention in the previous stage; Step S4-6: In the base training stage, only the original loss function needs to be calculated for the obtained output image; in the incremental training stage, a learnable perturbation is superimposed on the visual output image of the current stage model, and then the perturbation-superimposed image and the visual image output by the model in the previous incremental stage are respectively input into the dynamic discriminator network for adversarial learning, and then the output result of the discriminator is used to calculate the relative adversarial loss; Step S4-7: Optimize the entire network parameters by combining the losses calculated in steps S4-5 and S4-6; In step S4-5, the feature attention consistency loss is calculated using the learnable attention in the current stage and the frozen attention in the previous stage, which is specifically implemented as follows: First, the low-light training dataset of the current incremental stage t is in Represents low-light images, represents the corresponding normal light image, N t Represents the number of samples in the current domain, Input feature extractor separately and a frozen feature extractor The following formula is used to extract the intermediate hidden layer features: In the formula, L represents the number of feature hidden layers; indicates enhanced low-light pictures; represents the i-th hidden layer, i belongs to 1,...,L-1,L; refers to a nested function where The corresponding hidden layer features obtained by the feature extractor are: t ={f t 1 ,f t 2 ,…,f t L }in Introduce a learnable attention A for each incremental stage, and the attention of the current incremental stage is A t , the attention A of the previous incremental stage t-1 Guide the learning of attention in the current incremental stage, retain the important feature knowledge of the old domain, and use the feature attention consistency loss based on mean square error to optimize the model. The formula is as follows: In the formula, represents the feature attention consistency loss at the t-th incremental stage; In step S4-6, the relative adversarial loss is calculated using the output result of the discriminator, and the specific implementation method is as follows: First, the visual output of the old domain model is defined as the real sample, the visual output of the new domain model is defined as the generated sample, and a learnable perturbation R is superimposed on the visual output of the new domain model. t , designed to damage the image quality; then the generated samples and real samples with superimposed perturbations are input into the dynamic discriminator, and the output of the old domain model frozen in the previous incremental stage is expressed as: The output of the new domain model is represented as: Conduct adversarial training; In the incremental stage t, to optimize the generator, that is, the feature extractor The relative adversarial loss based on the least squares method is formulated as follows: In the formula, Represents the input data x t The expected value of p(x t ) represents the input data x t The true probability distribution of In the incremental stage t, to optimize the discriminator The relative adversarial loss based on the least squares method is formulated as follows: In the formula, the generator Try to minimize the objective function The adversarial discriminator Try to maximize it.

2. The robust incremental low-light image enhancement method according to claim 1, characterized in that: In step S1, the training order of different domain data sets in the incremental sample library is determined as follows: In the incremental learning setting, a domain dataset is randomly selected from the incremental sample library as the base dataset for the first training, and the remaining domain datasets are used as incremental domains. A single domain dataset is sequentially selected from the remaining incremental sample library as the domain dataset for the next training.

3. The robust incremental low-light image enhancement method according to claim 1, characterized in that: In step S3, the method for preprocessing the sample domain in real time is specifically as follows: Step S3-1: randomly flipping the low-light image horizontally and vertically with a preset probability; Step S3-2: Crop the low-light image to a uniform size; Step S3-3: convert the low-light image into a tensor, and convert the data type into float; Step S3-4: normalize each pixel of the low-light image so that the pixel value falls within the range of [-1, 1].

4. The robust incremental low-light image enhancement method according to claim 1, characterized in that: In step S4-7, in order to optimize the backbone network, i.e., the feature extractor Total loss The following formula: In the formula, represents the original loss function of the backbone network in the tth incremental stage, γ and δ are two hyperparameters, represents the feature attention consistency loss at the t-th incremental stage, represents the relative adversarial loss based on the least squares method at the incremental stage t; To optimize the dynamics discriminator Network, total loss The following formula: In the formula, η is a hyperparameter, is the optimized discriminator in the incremental stage t A least squares based relative adversarial loss is used.

5. The robust incremental low-light image enhancement method according to claim 1, characterized in that: In step S6, the model is comprehensively tested using a test data set to evaluate the enhancement effect and the forgetting rate. After the current domain model is trained, the model parameters obtained by the training are used to initialize the model parameters of the next incremental stage.

6. A robust incremental low-light image enhancement apparatus, comprising a processor and a computer program stored in a memory and executable on the processor, characterized in that: When the processor executes the program, the method according to any one of claims 1 to 5 is implemented.

Images